Resources — high-trust sources for Production AI Architecture

Ground every non-obvious claim in these; cite inline (<sup> → #sources). Populated before teaching. Prefer primary specs and vendor engineering docs over secondary commentary. This track is about running AI as a production system, so the spine is classic SRE / distributed-systems material with the AI-specific layers (telemetry conventions, risk framework, cloud lens) bolted on.

Primary specs / docs

• Google — Site Reliability Engineering (the SRE Book) (sre.google/books). The reliability vocabulary this whole track reuses: SLI/SLO/error budgets, graceful degradation, load shedding, toil, postmortems. AI doesn’t get a new reliability theory — it inherits this one.
• AWS — Well-Architected Framework: Generative AI Lens (docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens). Reviews a GenAI workload across the six pillars (op-excellence, security, reliability, performance, cost, sustainability) and ships reference scenarios — including a multi-tenant generative AI service, which maps directly to L5.
• OpenTelemetry — GenAI semantic conventions (opentelemetry.io/docs/specs/semconv/gen-ai/). The standard attribute set for LLM/agent spans: model name, input/output token counts, tool calls, tool results. The “trace/span” half of L2’s observability triad; treat it as the wire format so your telemetry isn’t vendor-locked.
• NIST — AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, 2023) (nist.gov/itl/ ai-risk-management-framework). The GOVERN / MAP / MEASURE / MANAGE functions — the backbone of L2’s governance + auditability material. See also the Generative AI Profile (NIST AI 600-1, 2024).
• OWASP — Top 10 for LLM Applications (owasp.org). Production threat model: prompt injection, sensitive-info disclosure, excessive agency. Feeds the security cut of the review methodology.

Papers & deeper reading

• Kleppmann, M., 2017 — “Designing Data-Intensive Applications” (DDIA). The reliability / scalability / maintainability framing and the failure-mode discipline (partial failure, retries, back-pressure, idempotency) that L1 ports onto an AI runtime. Not AI-specific; that’s the point.
• Beyer et al., 2016 — “Site Reliability Engineering” (O’Reilly / Google). Companion to the free SRE Book above for the error-budget and on-call operational model.
• Anthropic — “Building effective agents” (anthropic.com/engineering). Workflow-vs-agent boundary and where production risk concentrates in an agent runtime (feeds L4).

Notes on trust

• Cloud-vendor lenses describe best practices for that vendor’s stack (Bedrock/SageMaker for AWS); take the architecture shape as portable, treat product specifics as version-dependent.
• The OTel GenAI conventions are still evolving — capturing prompt/completion content is opt-in and carries PII/cost weight. Teach the attribute taxonomy, verify exact attribute names against the current spec before quoting them as fact.
• Cost and capacity numbers (tokens/sec, $/Mtok, GPU memory) drift fast — teach the model (queueing, batching, KV-cache memory), give concrete numbers only as illustrative “at time of writing.”